A2A Server: Agent-to-Agent Communication Framework

A lightweight, transport-agnostic framework for agent-to-agent communication based on JSON-RPC, implementing the A2A Protocol.

🚀 Quick Start

Install from PyPI

pip install a2a-server

Run with a Sample Agent

Create a minimal agent.yaml configuration file:

server:
  host: 0.0.0.0
  port: 8000

handlers:
  pirate_agent:
    type: a2a_server.tasks.handlers.google_adk_handler.GoogleADKHandler
    agent: a2a_server.sample_agents.pirate_agent.pirate_agent
    name: pirate_agent

Start the server:

uv run a2a-server --config agent.yaml

That's it! Your server is now running with a pirate-speaking agent.

🔍 What's in the Framework

A2A Server provides:

• Multiple Transport Layers: HTTP, WebSocket, Server-Sent Events (SSE)
• Flexible Handler System: Easily create and register custom agent handlers
• Google ADK Integration: Seamless use of Google Agent Development Kit agents
• Auto-Discovery: Find handlers through packages or entry points
• Agent Cards: A2A Protocol compatible agent descriptions
• Metrics & Observability: OpenTelemetry and Prometheus support
• Async-First Design: Modern concurrency with asyncio.TaskGroup

🤖 Using the Built-in Agents

A2A Server comes with sample agents that you can use right away:

• Pirate Agent: Converts text into pirate-speak
• Chef Agent: Provides cooking advice and recipes
• Echo Agent: Simple agent that echoes messages (useful for testing)

Configure Multiple Agents

Update your agent.yaml to include multiple agents:

server:
  host: 0.0.0.0
  port: 8000

handlers:
  use_discovery: false
  default: pirate_agent  # This will be the default handler

  pirate_agent:
    type: a2a_server.tasks.handlers.google_adk_handler.GoogleADKHandler
    agent: a2a_server.sample_agents.pirate_agent.pirate_agent
    name: pirate_agent

  chef_agent:
    type: a2a_server.tasks.handlers.google_adk_handler.GoogleADKHandler
    agent: a2a_server.sample_agents.chef_agent.chef_agent
    name: chef_agent

🛠️ Creating Your Own Agent

1. Create a Google ADK Agent

# my_agents/trivia_agent.py
from google.adk.agents import Agent
from google.adk.models.lite_llm import LiteLlm

trivia_agent = Agent(
    name="trivia_agent",
    model=LiteLlm(model="openai/gpt-4o-mini"),
    description="Provides fun trivia facts",
    instruction="You are a trivia expert. When users ask you questions, provide interesting and accurate trivia facts related to their query. Keep your responses brief, entertaining, and educational."
)

2. Register in Your Config

# agent.yaml
handlers:
  trivia_agent:
    type: a2a_server.tasks.handlers.google_adk_handler.GoogleADKHandler
    agent: my_agents.trivia_agent.trivia_agent
    name: trivia_agent

3. Start the Server with Your Module in Python Path

PYTHONPATH=/path/to/my_agents_directory a2a-server --config agent.yaml

🧪 Creating a Custom Handler

For more advanced use cases, create a custom handler:

# my_handlers/custom_handler.py
import asyncio
from a2a_server.tasks.handlers.task_handler import TaskHandler
from a2a_json_rpc.spec import (
    Message, TaskStatus, TaskState, Artifact, TextPart,
    TaskStatusUpdateEvent, TaskArtifactUpdateEvent
)

class CustomHandler(TaskHandler):
    @property
    def name(self) -> str:
        return "custom"
    
    async def process_task(self, task_id, message, session_id=None):
        # First yield a "working" status
        yield TaskStatusUpdateEvent(
            id=task_id,
            status=TaskStatus(state=TaskState.working),
            final=False
        )
        
        # Extract text from message
        text = ""
        if message.parts:
            part_data = message.parts[0].model_dump(exclude_none=True)
            if "text" in part_data:
                text = part_data["text"] or ""
        
        # Process the message (your custom logic here)
        response_text = f"Custom response to: {text}"
        
        # Create and yield an artifact
        response_part = TextPart(type="text", text=response_text)
        artifact = Artifact(name="custom_response", parts=[response_part], index=0)
        yield TaskArtifactUpdateEvent(id=task_id, artifact=artifact)
        
        # Finally, yield completion status
        yield TaskStatusUpdateEvent(
            id=task_id,
            status=TaskStatus(state=TaskState.completed),
            final=True
        )

Register it in your agent.yaml:

handlers:
  custom:
    type: my_handlers.custom_handler.CustomHandler
    name: custom

🧪 Testing Your Agents

Using a2a-cli

The easiest way to test your agents is with a2a-cli:

# Install the CLI
pip install a2a-cli

# Connect to your server's default agent
a2a-cli --server http://localhost:8000

# Connect to a specific agent
a2a-cli --server http://localhost:8000/pirate_agent

Using curl

Create a Task

curl -X POST http://localhost:8000/rpc \
  -H 'Content-Type: application/json' \
  -d '{
    "jsonrpc":"2.0",
    "id":1,
    "method":"tasks/send",
    "params":{
      "message":{
        "role":"user",
        "parts":[{"type":"text","text":"Tell me about pirates"}]
      }
    }
  }'

Stream Events for a Task

curl -N http://localhost:8000/events?task_ids=<task_id>

Use a Specific Agent

curl -X POST http://localhost:8000/chef_agent/rpc \
  -H 'Content-Type: application/json' \
  -d '{
    "jsonrpc":"2.0",
    "id":1,
    "method":"tasks/send",
    "params":{
      "message":{
        "role":"user",
        "parts":[{"type":"text","text":"What can I make with chicken and rice?"}]
      }
    }
  }'

Get Agent Metadata

curl http://localhost:8000/.well-known/agent.json
curl http://localhost:8000/pirate_agent/.well-known/agent.json

Using WebSockets

For WebSocket testing, use the websocat tool:

# Install websocat
cargo install websocat

# Connect to default agent
websocat ws://localhost:8000/ws

# Connect to specific agent
websocat ws://localhost:8000/pirate_agent/ws

Then send a JSON-RPC request:

{"jsonrpc":"2.0","id":1,"method":"tasks/send","params":{"message":{"role":"user","parts":[{"type":"text","text":"Hello there!"}]}}}

🚀 Advanced Configuration

Enable Metrics

# Enable Prometheus metrics
export PROMETHEUS_METRICS=true
a2a-server --config agent.yaml

# Use OpenTelemetry
export OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4318
a2a-server --config agent.yaml

Log Levels

# Set log level
a2a-server --config agent.yaml --log-level debug

Handler Discovery

# Add custom handler packages
a2a-server --config agent.yaml --handler-package my_custom_handlers

📚 API Endpoints

Endpoint	Description
/rpc	JSON-RPC endpoint for the default handler
/{handler}/rpc	JSON-RPC endpoint for a specific handler
/ws	WebSocket endpoint for the default handler
/{handler}/ws	WebSocket endpoint for a specific handler
/events	SSE endpoint for the default handler
/{handler}/events	SSE endpoint for a specific handler
/.well-known/agent.json	Agent Card for the default handler
/{handler}/.well-known/agent.json	Agent Card for a specific handler
/metrics	Prometheus metrics (when enabled)

🔒 Security Note

For production deployment, we recommend placing the A2A server behind an authentication layer, as the SSE and WebSocket endpoints do not include authentication by default.

🌟 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.